Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias

Zixiang Xu, Sixian Li, Huaxing Liu, Xiang Wang, Shuai Li et al.|July 13, 2026arXiv

Key Takeaway

LLM judge bias lives in a specific geometric structure within the model's activations, making it detectable and controllable without changing prompts—useful for building more reliable evaluation systems.

Summary

This paper reveals that LLM judge bias isn't just about input-output relationships—it's encoded in the model's internal representations. The researchers found that biased inputs activate a consistent, low-dimensional subspace in the model's hidden states, which they can steer to control bias direction and predict when judges will fail on new tasks.

evaluation safety

Key Terms

activation-manifold steering-vectors llm-as-judge mechanistic-interpretability evaluator-bias